Memory Map
shandar/memory-map
Memória persistente para Claude entre sessões com snapshot, recall e limpeza
Ver no GitHubMCP server que cria um loop completo de memória para Claude, permitindo que a IA mantenha contexto de projeto entre sessões. Implementa três skills: snapshot (captura estado), recall (recupera memória) e prune (limpa dados obsoletos). Funciona em claude.ai, Claude Code, Cursor e Windsurf.
Instalação
npx convez-ai shandar/memory-map1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Manter contexto de projetos longos sem repetir explicações a cada sessão
- Rastrear decisões, componentes construídos e tarefas pendentes automaticamente
- Recuperar estado de projeto no início de nova conversa sem reescrever tudo
Veredito

Solução prática que resolve um pain point real do fluxo Claude—sessões começam do zero. A implementação é técnica (shell/SKILL.md), bem documentada e funciona em múltiplos ambientes. Recomendado para quem trabalha em projetos contínuos.
README
A tradução para o português deste README ainda está na fila.
memory-map
A complete memory loop for Claude. Snapshot, recall, and prune project context across sessions.
Three skills, bundled as one plugin. Works in claude.ai, Claude Code, Cursor, Windsurf, and Codex CLI.
The problem
Every new Claude conversation starts blank.
You re-explain your project. You re-describe what you are building. You re-state the decisions you already made. Every. Single. Time.
It is a tax on every session. And it gets worse the longer your project runs.
The fix
Three skills that give Claude a working memory across sessions:
| Skill | What it does | When it triggers |
|---|---|---|
| memory-map | Snapshots your session — what was built, decided, pending — to persistent memory | At session close, or when you say "save this" |
| memory-recall | Loads stored memory and surfaces what is relevant right now | At session start, or when you ask "what do you remember?" |
| memory-prune | Identifies stale entries and proposes a cleanup, with your confirmation | When memory gets cluttered, or on request |
The loop: write → recall → prune. Together they make Claude feel like it actually remembers your work.
Works everywhere
Same protocol, two storage backends, detected automatically:
| Environment | Storage |
|---|---|
| claude.ai | Persistent memory API (memory_user_edits) |
| Claude Code | Local file at ~/.claude/memory.md |
Compatible with the open SKILL.md format, also works in Cursor, Windsurf, Codex CLI, and any tool that supports Anthropic's Agent Skills spec.
Install
Claude Code (one line)
curl -fsSL https://raw.githubusercontent.com/shandar/memory-map/main/install.sh | bash
The installer clones the repo, copies all three skills to ~/.claude/skills/, initialises ~/.claude/memory.md, and backs up any pre-existing skill folders.
Claude Code (manual)
git clone https://github.com/shandar/memory-map.git
cd memory-map
cp -r skills/* ~/.claude/skills/
Restart Claude Code. The skills load automatically.
claude.ai
Each skill installs separately on claude.ai:
- Download the latest release
- Extract the archive
- Open Settings → Skills → Install from file
- Upload
skills/memory-map/SKILL.md, thenmemory-recall/SKILL.md, thenmemory-prune/SKILL.md
After all three are installed, they trigger automatically based on what you say.
Usage
After install, just talk normally. The skills trigger themselves.
At the start of a session
You: "What do you remember about my dashboard project?"
Claude: Picking up where we left off:
Current state — Auth flow shipped to staging. API rate limiting
implemented with Redis backend. Dashboard refactor in progress.
Pending items:
→ Finish the analytics widget refactor
→ Review PR #142 from Sarah
→ Migrate billing service to new endpoint
Active projects: dashboard refactor, auth flow, billing migration.
At the end of a session
You: "Save this conversation"
Claude: MEMORY MAP — updated April 27 2026
UPDATED → Auth flow status (now in staging)
UPDATED → Current state snapshot
NEW → API rate limiting decisions (Redis backend)
──────────────────────────────────────────────────
UNCHANGED: 8 entries
When memory gets cluttered
You: "Clean up the memory"
Claude: MEMORY PRUNE — proposed changes
REMOVE (stale)
→ Entry 7: "Old auth design (v1)" — superseded by v3 in entry 11
→ Entry 14: "Q1 sprint planning" — quarter is closed
CONSOLIDATE
→ Entries 3 + 9 → single "API conventions" entry
KEEP all 6 operating preference entries
Apply these changes? [yes / modify / cancel]
Trigger phrases
The skills activate on natural language, no commands to memorise.
| Skill | Trigger phrases |
|---|---|
| memory-map | save this · snapshot this · memorise this chat · update your memory · store this session · don't lose this context |
| memory-recall | what do you remember · what's the context · where did we leave off · catch me up · show me the memory map |
| memory-prune | clean up memory · prune memory · consolidate memory · trim the memory map |
All three also fire proactively when conditions are right — at the close of a productive session, at the start of a new one, or when memory is approaching its limits.
How memory is stored
Claude Code: ~/.claude/memory.md
Plain markdown, readable, editable, version-controllable:
# Claude Memory Map
Last updated: April 27 2026
## 1. CURRENT STATE SNAPSHOT (April 2026)
Auth flow in staging. API rate limiting shipped with Redis. Dashboard refactor
in progress. Next: analytics widget, PR #142 review, billing migration.
## 2. AUTH SERVICE
JWT with refresh tokens. Redis session store. 15-min access, 7-day refresh.
PR: github.com/team/repo/pull/127. Owner: backend team.
## 3. CODING CONVENTIONS
TypeScript strict mode. Vitest for tests. Conventional commits. Trunk-based dev.
You can edit this file directly. memory-prune will respect your edits.
claude.ai: Persistent memory API
Stored via Anthropic's memory_user_edits tool. Visible in Settings → Memory. Works the same way functionally — same category labels, same density rules, just managed via API instead of file.
Design principles
- Proactive, not reactive — skills run when they should, without being asked
- Replace over accumulate — stale entries are replaced, never stacked
- Density over prose — every entry is scannable, fact-packed, no filler
- One source of truth —
CURRENT STATE SNAPSHOTis always exactly one entry - Confirm before destruction — memory-prune always proposes, never deletes autonomously
- Environment-aware — same skill, right storage for the right tool
Repository structure
memory-map/
├── README.md ← you are here
├── LICENSE ← MIT
├── CHANGELOG.md ← version history
├── CONTRIBUTING.md ← how to contribute
├── marketplace.json ← plugin manifest
├── install.sh ← one-line installer
└── skills/
├── memory-map/
│ ├── SKILL.md
│ └── references/
│ ├── claude-ai.md ← memory_user_edits API
│ └── claude-code.md ← ~/.claude/memory.md file
├── memory-recall/
│ ├── SKILL.md
│ └── references/
│ ├── claude-ai.md
│ └── claude-code.md
└── memory-prune/
├── SKILL.md
└── references/
├── claude-ai.md
└── claude-code.md
Each skill is self-contained and can be used independently. The plugin bundles them for one-command install.
Roadmap
- v0.1 — memory-map, memory-recall, memory-prune
- v0.2 — multi-project memory (separate stores per repo or workspace)
- v0.3 — memory export/import for sharing project context across teams
- v0.4 — memory diffing across sessions ("what changed since last week?")
Track progress in issues and discussions.
Built by
Shandar Junaid — Founder, Affordance Design Studio
If you build on top of this, I would love to see what you make. Tag me on LinkedIn or X.
License
MIT — use it, fork it, extend it. See LICENSE.
Credits
The SKILL.md format and Agent Skills specification are from Anthropic. This plugin is an independent open-source project, not affiliated with Anthropic.